ppc.pl

mediastreamer2是开源的网络传输媒体流的库

	$LD		r6,`1*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6	
	
	addc		r10,r7,r10		#add the two register number
	adde		r11,r8,r0 		# (r8,r7) to the three register
	addze		r9,r0			# number (r9,r11,r10).NOTE:r0=0
	
	addc		r10,r7,r10		#add the two register number
	adde		r11,r8,r11 		# (r8,r7) to the three register
	addze		r9,r9			# number (r9,r11,r10).
	
	$ST		r10,`1*$BNSZ`(r3)	# r[1]=c2
				
						#sqr_add_c(a,1,c3,c1,c2);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r0
						#sqr_add_c2(a,2,0,c3,c1,c2);
	$LD		r6,`2*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	
	$ST		r11,`2*$BNSZ`(r3)	#r[2]=c3
						#sqr_add_c2(a,3,0,c1,c2,c3);
	$LD		r6,`3*$BNSZ`(r4)	#r6 = a[3]. r5 is already a[0].
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r0
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
						#sqr_add_c2(a,2,1,c1,c2,c3);
	$LD		r5,`1*$BNSZ`(r4)
	$LD		r6,`2*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	
	$ST		r9,`3*$BNSZ`(r3)	#r[3]=c1;
						#sqr_add_c(a,2,c2,c3,c1);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r0
						#sqr_add_c2(a,3,1,c2,c3,c1);
	$LD		r6,`3*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
						#sqr_add_c2(a,4,0,c2,c3,c1);
	$LD		r5,`0*$BNSZ`(r4)
	$LD		r6,`4*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	$ST		r10,`4*$BNSZ`(r3)	#r[4]=c2;
						#sqr_add_c2(a,5,0,c3,c1,c2);
	$LD		r6,`5*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r0
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
						#sqr_add_c2(a,4,1,c3,c1,c2);
	$LD		r5,`1*$BNSZ`(r4)
	$LD		r6,`4*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
						#sqr_add_c2(a,3,2,c3,c1,c2);
	$LD		r5,`2*$BNSZ`(r4)
	$LD		r6,`3*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	$ST		r11,`5*$BNSZ`(r3)	#r[5]=c3;
						#sqr_add_c(a,3,c1,c2,c3);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r0
						#sqr_add_c2(a,4,2,c1,c2,c3);
	$LD		r6,`4*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
						#sqr_add_c2(a,5,1,c1,c2,c3);
	$LD		r5,`1*$BNSZ`(r4)
	$LD		r6,`5*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
						#sqr_add_c2(a,6,0,c1,c2,c3);
	$LD		r5,`0*$BNSZ`(r4)
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	$ST		r9,`6*$BNSZ`(r3)	#r[6]=c1;
						#sqr_add_c2(a,7,0,c2,c3,c1);
	$LD		r6,`7*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r0
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
						#sqr_add_c2(a,6,1,c2,c3,c1);
	$LD		r5,`1*$BNSZ`(r4)
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
						#sqr_add_c2(a,5,2,c2,c3,c1);
	$LD		r5,`2*$BNSZ`(r4)
	$LD		r6,`5*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
						#sqr_add_c2(a,4,3,c2,c3,c1);
	$LD		r5,`3*$BNSZ`(r4)
	$LD		r6,`4*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	$ST		r10,`7*$BNSZ`(r3)	#r[7]=c2;
						#sqr_add_c(a,4,c3,c1,c2);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r0
						#sqr_add_c2(a,5,3,c3,c1,c2);
	$LD		r6,`5*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
						#sqr_add_c2(a,6,2,c3,c1,c2);
	$LD		r5,`2*$BNSZ`(r4)
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
						#sqr_add_c2(a,7,1,c3,c1,c2);
	$LD		r5,`1*$BNSZ`(r4)
	$LD		r6,`7*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	$ST		r11,`8*$BNSZ`(r3)	#r[8]=c3;
						#sqr_add_c2(a,7,2,c1,c2,c3);
	$LD		r5,`2*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r0
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
						#sqr_add_c2(a,6,3,c1,c2,c3);
	$LD		r5,`3*$BNSZ`(r4)
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
						#sqr_add_c2(a,5,4,c1,c2,c3);
	$LD		r5,`4*$BNSZ`(r4)
	$LD		r6,`5*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	$ST		r9,`9*$BNSZ`(r3)	#r[9]=c1;
						#sqr_add_c(a,5,c2,c3,c1);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r0
						#sqr_add_c2(a,6,4,c2,c3,c1);
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
						#sqr_add_c2(a,7,3,c2,c3,c1);
	$LD		r5,`3*$BNSZ`(r4)
	$LD		r6,`7*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	$ST		r10,`10*$BNSZ`(r3)	#r[10]=c2;
						#sqr_add_c2(a,7,4,c3,c1,c2);
	$LD		r5,`4*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r0
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
						#sqr_add_c2(a,6,5,c3,c1,c2);
	$LD		r5,`5*$BNSZ`(r4)
	$LD		r6,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	addc		r11,r7,r11
	adde		r9,r8,r9
	addze		r10,r10
	$ST		r11,`11*$BNSZ`(r3)	#r[11]=c3;
						#sqr_add_c(a,6,c1,c2,c3);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r0
						#sqr_add_c2(a,7,5,c1,c2,c3)
	$LD		r6,`7*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	addc		r9,r7,r9
	adde		r10,r8,r10
	addze		r11,r11
	$ST		r9,`12*$BNSZ`(r3)	#r[12]=c1;
	
						#sqr_add_c2(a,7,6,c2,c3,c1)
	$LD		r5,`6*$BNSZ`(r4)
	$UMULL		r7,r5,r6
	$UMULH		r8,r5,r6
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r0
	addc		r10,r7,r10
	adde		r11,r8,r11
	addze		r9,r9
	$ST		r10,`13*$BNSZ`(r3)	#r[13]=c2;
						#sqr_add_c(a,7,c3,c1,c2);
	$UMULL		r7,r6,r6
	$UMULH		r8,r6,r6
	addc		r11,r7,r11
	adde		r9,r8,r9
	$ST		r11,`14*$BNSZ`(r3)	#r[14]=c3;
	$ST		r9, `15*$BNSZ`(r3)	#r[15]=c1;


	bclr	BO_ALWAYS,CR0_LT

	.long	0x00000000

#
#	NOTE:	The following label name should be changed to
#		"bn_mul_comba4" i.e. remove the first dot
#		for the gcc compiler. This should be automatically
#		done in the build
#

.align	4
.bn_mul_comba4:
#
# This is an optimized version of the bn_mul_comba4 routine.
#
# void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
# r3 contains r
# r4 contains a
# r5 contains b
# r6, r7 are the 2 BN_ULONGs being multiplied.
# r8, r9 are the results of the 32x32 giving 64 multiply.
# r10, r11, r12 are the equivalents of c1, c2, and c3.
#
	xor	r0,r0,r0		#r0=0. Used in addze below.
					#mul_add_c(a[0],b[0],c1,c2,c3);
	$LD	r6,`0*$BNSZ`(r4)		
	$LD	r7,`0*$BNSZ`(r5)		
	$UMULL	r10,r6,r7		
	$UMULH	r11,r6,r7		
	$ST	r10,`0*$BNSZ`(r3)	#r[0]=c1
					#mul_add_c(a[0],b[1],c2,c3,c1);
	$LD	r7,`1*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r11,r8,r11
	adde	r12,r9,r0
	addze	r10,r0
					#mul_add_c(a[1],b[0],c2,c3,c1);
	$LD	r6, `1*$BNSZ`(r4)		
	$LD	r7, `0*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r11,r8,r11
	adde	r12,r9,r12
	addze	r10,r10
	$ST	r11,`1*$BNSZ`(r3)	#r[1]=c2
					#mul_add_c(a[2],b[0],c3,c1,c2);
	$LD	r6,`2*$BNSZ`(r4)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r12,r8,r12
	adde	r10,r9,r10
	addze	r11,r0
					#mul_add_c(a[1],b[1],c3,c1,c2);
	$LD	r6,`1*$BNSZ`(r4)		
	$LD	r7,`1*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r12,r8,r12
	adde	r10,r9,r10
	addze	r11,r11
					#mul_add_c(a[0],b[2],c3,c1,c2);
	$LD	r6,`0*$BNSZ`(r4)		
	$LD	r7,`2*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r12,r8,r12
	adde	r10,r9,r10
	addze	r11,r11
	$ST	r12,`2*$BNSZ`(r3)	#r[2]=c3
					#mul_add_c(a[0],b[3],c1,c2,c3);
	$LD	r7,`3*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r10,r8,r10
	adde	r11,r9,r11
	addze	r12,r0
					#mul_add_c(a[1],b[2],c1,c2,c3);
	$LD	r6,`1*$BNSZ`(r4)
	$LD	r7,`2*$BNSZ`(r5)
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r10,r8,r10
	adde	r11,r9,r11
	addze	r12,r12
					#mul_add_c(a[2],b[1],c1,c2,c3);
	$LD	r6,`2*$BNSZ`(r4)
	$LD	r7,`1*$BNSZ`(r5)
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r10,r8,r10
	adde	r11,r9,r11
	addze	r12,r12
					#mul_add_c(a[3],b[0],c1,c2,c3);
	$LD	r6,`3*$BNSZ`(r4)
	$LD	r7,`0*$BNSZ`(r5)
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r10,r8,r10
	adde	r11,r9,r11
	addze	r12,r12
	$ST	r10,`3*$BNSZ`(r3)	#r[3]=c1
					#mul_add_c(a[3],b[1],c2,c3,c1);
	$LD	r7,`1*$BNSZ`(r5)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r11,r8,r11
	adde	r12,r9,r12
	addze	r10,r0
					#mul_add_c(a[2],b[2],c2,c3,c1);
	$LD	r6,`2*$BNSZ`(r4)
	$LD	r7,`2*$BNSZ`(r5)
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r11,r8,r11
	adde	r12,r9,r12
	addze	r10,r10
					#mul_add_c(a[1],b[3],c2,c3,c1);
	$LD	r6,`1*$BNSZ`(r4)
	$LD	r7,`3*$BNSZ`(r5)
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r11,r8,r11
	adde	r12,r9,r12
	addze	r10,r10
	$ST	r11,`4*$BNSZ`(r3)	#r[4]=c2
					#mul_add_c(a[2],b[3],c3,c1,c2);
	$LD	r6,`2*$BNSZ`(r4)		
	$UMULL	r8,r6,r7
	$UMULH	r9,r6,r7
	addc	r12,r8,r12
	adde	r10,r9,r10
	addze	r11,r0
					#mul_add_c(a[3],b[2],c3,c1,c2);